Validation of AI algorithms in ophthalmological population screening

Abstract

The application of artificial intelligence (AI) in ophthalmology has emerged as a promising strategy for the early diagnosis and population screening of visual diseases, mitigating the shortage of specialists and optimizing the management of health resources. The present study aims to analyze the clinical efficacy, robustness, and generalization capacity of AI algorithms applied to the screening and diagnostic support of ophthalmological pathologies. For this purpose, an integrative literature review was conducted through searches in the Google Scholar, PubMed, ResearchGate, and CAPES Periodicals Portal databases, culminating in the selection and analysis of ten studies published between 2018 and 2025. The synthesis of evidence demonstrated that deep learning models focused on the analysis of fundus images present high rates of sensitivity and specificity, especially in the detection of diabetic retinopathy, retinopathy of prematurity, glaucoma, and age-related macular degeneration. Furthermore, it was found that the integration of these technologies significantly expands clinical screening capacity. However, the performance of these algorithms can undergo significant variations when faced with distinct populations or different capture equipment, a technical challenge known as domain shift. Therefore, it is inferred that the safe clinical implementation of AI requires rigorous multicenter external validations and the development of clear ethical-regulatory guidelines, acting as a complement to medical judgment and democratizing access to visual care.